Events
DMS Statistics and Data Science Seminar |
| Time: Feb 11, 2026 (02:00 PM) |
| Location: 358 Parker Hall |
|
Details:
Speaker: Sayar Karmakar (University of Florida, Department of Statistics) Title: Epidemic Changepoints: Applications in spatial anomaly detection and localizing LLM watermarks
Abstract: We present epidemic change-points as a unifying lens for two localization problems: (i) detecting spatial anomalies and (ii) segmenting watermarked regions in mixed-source text. For spatial data, we formalize a "spatial" change-point as an anomalous region (an epidemic in space), provide detection-accuracy results for single and multiple breaks, and propose a block-based scan that delivers substantial computational savings with guarantees. Next, we move to a seemingly unrelated but a very pertinent topic. As large language models proliferate, ensuring content provenance has become a statistical challenge. For this problem on finding localized modified text data segments, we introduce WISER, a fast epidemic-segmentation approach with finite-sample error bounds and consistency for multiple watermarked segments, and we demonstrate empirical gains over state-of-the-art baselines on benchmark datasets.
We emphasize how classical changepoint ideas catered to epidemic and transient departures yield principled, scalable solutions to modern problems in text provenance and spatial anomaly detection. Simulations and empirical studies corroborate the theory and point to open questions for PhD-level research.
Joint work with Soham Bonnerjee and Subhrajyoty Roy (watermarks) and with Soham Bonnerjee and George Michailidis (spatial anomaly).
Speaker Bio: Sayar Karmakar is an Assistant Professor of Statistics at the University of Florida. His research spans high-dimensional time series, changepoints, spatial and spatiotemporal data, econometrics and applied probability.
|
